# Stochastic optimal control to a nonlinear differential game

- Othusitse Basimanebotlhe
^{1, 2}Email author and - Xiaoping Xue
^{1}

**2014**:266

https://doi.org/10.1186/1687-1847-2014-266

© Basimanebotlhe and Xue; licensee Springer. 2014

**Received: **9 April 2014

**Accepted: **24 September 2014

**Published: **14 October 2014

## Abstract

The paper studies the optimal control of a nonlinear stochastic differential game of two persons subjected to noisy measurements. The logarithmic transformation to the value function is used in trying to find the solution of the problem. The conversion of a quasilinear partial differential equation to an ordinary linear differential equation is considered. Lastly, the iterative optimal control path estimates for the minimization maximization differential game are attained.

## Keywords

## 1 Introduction

Control theory is a field of mathematics and engineering used in a wide range of fields and their applications, such as architecture, communications, queueing theory, robotics and in economics as evidenced in [1–3] just to mention a few. Control theory is a subject of much interest in today’s real world. As stated in [4], optimizing a sequence of actions to attain some future goal is the general topic of control theory. Therefore, the objective of optimal control theory is to attain an optimal regulation of the system evolution [5]. Without the indulgence of the noise, the continuous time control problems can be solved in two ways: using Pontryagin Minimum Principle (PMP), which is a pair of ordinary differential equations, or the Hamilton-Jacobi-Bellman (HJB), which is a partial differential equation as in [4]. The addition of differential equations as constraints in the optimization problem leads to the property that in optimal control theory the minimum is no longer represented by one point ${x}^{\ast}$ in the state space but by a path or trajectory ${x}^{\ast}={({x}_{i}^{\ast})}_{i=1,\dots ,N}$, which is known as the optimal trajectory.

In the presence of noise, the PMP formalism has no obvious generalization as mentioned in [6]. However, the inclusion of the noise in the HJB framework is mathematically quite straightforward, while the numerical solution of either deterministic or stochastic HJB equation is difficult due to the curse of dimensionality. A control problem is said to be stochastic when it is subjected to some disturbances or noise terms and time dependent, that is being uncertain of its future state. Under control theory, there lies a topic of interest to this paper which is game theory.

Game theory deals with strategic interactions among several decision makers, known as players. These players have objectives that may be contradicting or non-contradicting. In situations whereby players’ aims are not in contradiction, they are said to be in cooperation, which is enough to regard their aims as of one player, as seen in [7] and [8]. If the players’ decisions are in contradiction, then it is widely known as non-cooperative game theory that is even more interesting as each player may have to be analyzed individually or in distinct groups basing on the similarity of the objectives. Non-cooperative game theory has been discussed by several authors [9–12].

The nonlinear stochastic optimal control theory is one of the optimization fundamentals with plethora of applications in several domains, see [13] and [14]. The main difficulty associated with stochastic minimax dynamic games studied here is the presence of the noise in the dynamical constraints and the solution to a nonlinear second order Hamilton Jacobi Bellman equation (HJB) as mentioned in [4].

In this paper a nonlinear stochastic problem is considered as a two-person zero sum game. One player tries to attain the reward at minimal costs, while the other player tries to maximize the costs of the other player, knowing that it would be advantageous to him. The problem is modeled as a dynamical system with nonlinear stochastic constraints, and our idea is to find the saddle point of the game after a certain time of play. The two controls act as opponents, that is, the control *u* is a stabilizer or minimizer, while *v* is the destabilizing or maximizing control variable. A different approach to finding the solution of the problem deployed through certain conditions assumed in this paper is similar to that of [15].

## 2 Problem formulation

Let $(\mathrm{\Omega},\mathcal{F},\mathbb{P},{\mathcal{F}}_{t})$ be a complete probability space, and let time interval $[0,T]$ with $0<T<\mathrm{\infty}$ be given. Assume that on this space an *n*-dimensional Brownian motion ${\{W(t),{\mathcal{F}}_{t}\}}_{t\in [0,T]}$ is defined with ${\{{\mathcal{F}}_{t}\}}_{t\in [0,T]}$ as the Brownian filtration, Guo *et al.* [16]. The expectation under the probability measure ℙ will be denoted by ${E}^{t}$, while ℱ is a *σ*-algebra of the subsets of Ω.

**Definition 2.1** (Brownian motion)

- (i)
At any time $s<t$, the increment $W(t)-W(s)$ is Gaussian with mean zero and variance ${E}^{t}[{(W(t)-W(s))}^{2}]=t-s$; moreover, the increments associated with disjoint intervals are independent.

- (ii)
Its sample path is continuous, that is, the function $t\to W(t)$ is almost surely continuous.

- (iii)
Its initial state and time is zero, that is, $W(0)=0$.

Let ${\{x(t,\omega )\}}_{t\in [0,T]}$ denote a stochastic process, then ${\{x(t,\omega )\}}_{t\in [0,T]}$ is said to be ${\mathcal{F}}_{t}$ adapted whenever ${\mathcal{F}}_{t}$ is known at time *t*.

*σ*is only dependent on the state trajectory

*x*, similarly as addressed by [18]. The function

*g*is dependent on the controls or policies, the state trajectory and the time

*t*. The value ${x}_{0}(\omega )$ denotes the initial random state. The problem is formulated as a differential game with two opponents as in Theodorou [15] with

- (i)
the function $\mathcal{L}:[0,T]\times \times {\mathbb{R}}^{n}\times {E}^{p}\times {E}^{q}\to \mathbb{R}$ is known as the immediate cost to go, while $\varphi :{\mathbb{R}}^{n}\to \mathbb{R}$ is the terminal cost;

- (ii)
$x(t,\omega )\in {\mathbb{L}}_{2}(\mathrm{\Omega},\mathcal{F},\mathbb{P})$ is the

*n*-dimensional random state vector for $t\ge 0$ with $\omega \in \mathrm{\Omega}$ as the supporting set of a complete probability measure space ℙ; - (iii)
$u(x(t,\omega ))\in {E}^{p}$ and $v(x(t,\omega ))\in {E}^{q}$ are the

*p*-dimensional and*q*-dimensional random control variables of player one and player two defined respectively in the two metric spaces; - (iv)
the symmetric and positive definite random time varying matrices $R(t,\omega )$ and $S(t,\omega )$ are matrices associated with the controls of $u(x(t,\omega ))$ and $v(x(t,\omega ))$, respectively;

- (v)
$\beta >0$ is the discounting factor of the value function.

**Definition 2.2** (Admissible control)

- (i)
for every $(t,x)$ the system of SDEs in (1) with initial condition $x(0,\omega )={x}_{0}(\omega )$ admits a pathwise unique strong solution;

- (ii)
there exists some function $\varphi :{\mathbb{R}}^{n}\to U$ of class ${C}^{1,2}$ such that

*u*is in relative feedback to*ϕ*, that is, $u(t)=\varphi (x(t))$ for every $t\in [0,T]$.

The two controls $u(x(t,\omega ))$ and $v(x(t,\omega ))$ are referred to as the stabilizing and the destabilizing controllers, respectively. The stabilizing controller minimizes the cost function, while the destabilizing controller tries to maximize the cost function.

- (i)
$f:[0,T]\times {\mathbb{R}}^{n}\to \mathbb{R}$ and $\sigma :[0,T]\times {\mathbb{R}}^{n}\to {\mathbb{R}}^{n\times m}$ be bounded and continuous functions;

- (ii)
$G(\omega )$ and $H(\omega )$ be $n\times p$ and $n\times q$ random matrices.

## 3 Conditions

- (i)
The functions $f(x)$, $L(t,x,u,v)$ are continuously differentiable in $(t,x,u,v)\in [0,T]\times {\mathbb{R}}^{n}\times {E}^{p}\times {E}^{q}$ and

*ϕ*is twice differentiable in $x\in {\mathbb{R}}^{n}$. - (ii)
*J*and ℒ are nonnegative functions. - (iii)
$V(x)$, ${\mathrm{\nabla}}_{x}V$, ℒ, and

*ϕ*are bounded, where*V*is defined as the value of the game that each control tries to optimize from. - (iv)The controls are bounded in ${\mathbb{R}}^{m}$ and given in the spaces$\begin{array}{c}{L}_{\mathcal{F}}^{2}:=\{u:{\parallel u(x(t,\omega ))\parallel}_{2}\}<\mathrm{\infty},\hfill \\ {L}_{\mathcal{F}}^{\mathrm{\infty}}:=\{u:{\parallel u(x(t,\omega ))\parallel}_{\mathrm{\infty}}\}<\mathrm{\infty},\hfill \end{array}$

for *v* continuous. The same conditions for *u* under (iv) are applicable to *v* as well.

## 4 Approach to the solution of stochastic optimal controls

Our approach in finding the optimal control is based on the definition of the saddle point given below as in [20] with slight changes to suit our problem.

**Definition 4.1**

- (i)If the pair $({u}^{\ast}(x(t,\omega )),{v}^{\ast}(x(t,\omega )))\in {U}_{1}\times {U}_{2}$ is optimal, then there exists a saddle point of the game over the interval $[0,T]$ with respect to $x(t,\omega )\in {\mathbb{R}}^{n}$, if$\begin{array}{c}J(t,x(t,\omega ),{u}^{\ast}(x(t,\omega )),v(x(t,\omega )))\hfill \\ \phantom{\rule{1em}{0ex}}\le J(t,x(t,\omega ),{u}^{\ast}(x(t,\omega )),{v}^{\ast}(x(t,\omega )))\hfill \\ \phantom{\rule{1em}{0ex}}\le J(t,x(t,\omega ),u(x(t,\omega )),{v}^{\ast}(x(t,\omega )))\hfill \end{array}$

- (ii)The upper value of the game at any path $x(t,\omega )$ and time $t\in [0,T]$ is defined by${V}^{\ast}(x(t,\omega ))=\underset{u\in {U}_{1}}{inf}\underset{v\in {U}_{2}}{sup}J(t,x(t,\omega ),u(x(t,\omega )),v(x(t,\omega ))),$

The objective is to find the optimal admissible controls, ${u}^{\ast}(x(t,\omega ))\in {U}_{1}$ and ${v}^{\ast}(x(t,\omega ))\in {U}_{2}$, such that $V(x(t,\omega ),t)$ satisfies Definition 4.1 for ${U}_{1}\subset {E}^{p}$ and ${U}_{2}\subset {E}^{q}$.

**Theorem 4.1** (Bellman principle of optimality)

*If* ${u}^{\ast}(x(t,\omega ))$ *is optimal over the interval* $[0,T]$ *starting at an initial state* ${x}_{0}(\omega )$, *then* ${u}^{\ast}(x(t,\omega ))$ *is necessarily optimal over the subinterval* $[t,t+dt]$ *for any* *dt* *such that* $T-t\ge dt>0$.

For the proof of the above theorem, refer to [15].

- (i)
$H{S}^{-1}{H}^{T}-G{R}^{-1}{G}^{T}<0$ implies that more weight is on the minimizing control than on the maximizing control variable.

- (ii)
$H{S}^{-1}{H}^{T}-G{R}^{-1}{G}^{T}>0$ implies more weight on the maximizing control than on the minimizing control variable.

- (iii)
$H{S}^{-1}{H}^{T}-G{R}^{-1}{G}^{T}=0$, the weights of the controls are equivalent, hence it is an ideal situation for a minimax optimal control.

The intuition we get from [13] is that the higher the variance, the lower the weight of the controls, hence ‘cheap’ controls and *vice versa*. In our case we want to strike a deal such that both players attain their optimums. The variance of the Brownian noise here is given by $\sigma {\sigma}^{T}>0$, therefore we want to attain a situation whereby $\lambda (t)[G{R}^{-1}{G}^{T}-H{S}^{-1}{H}^{T}]=\sigma {\sigma}^{T}$ for all $x\in {\mathbb{R}}^{n}$ and $t\in [0,T]$, where the difference of the control coefficients will be the same as the variance of the noise. Our assumption on the balancing parameter is different from the one suggested by other authors, as in [13] and [15], where the balancing term is just a constant parameter. In our case, the balancing variable $\lambda (t)$ is dependent on *t* such that at any time instant the equality sign is attained as the variance terms differing with time.

If the solution $\mathrm{\Phi}(x(t,\omega ))$ is found to exist for equation (16), then we have the results given below.

**Theorem 4.2**

*If*$\mathrm{\Phi}(x(t,\omega ))$

*satisfies equation*(16),

*then the transformed control optimums are given as*

*and*

*for the value*

*where*$\lambda (t)$

*satisfies*

One would observe that ${u}^{\ast}(x(t,\omega ))$ is now positive while ${v}^{\ast}(x(t,\omega ))$, this is so because the problem has been transformed from minimax to maxmin problem. The PDE in (16) is found to be a bit difficult to solve in terms of dependence variables *x* and *t*, therefore in this paper we resort to transforming the above PDE to an ODE for which, in most cases, a solution can be obtained. Consider a one-dimensional problem for this case, thus $n=1$ and fix *t*, then the equation becomes more dependent on *x*. This leads to a nonlinear ODE, and before solving the nonlinear ODE, we have the following assumptions.

- (i)
$\mathrm{\Phi}(x(t,\omega ),t)$, $f(x(t,\omega ))$ and $q(x(t,\omega ))$ are nonnegative functions.

- (ii)
$\mathrm{\Phi}(x(t,\omega ),t)$ is Lipschitz continuous for all $(t,x)\in ([0,T]\times \mathbb{R})$ and $\omega \in \mathrm{\Omega}$.

- (iii)
$f(x(t,\omega ))$ and $q(x(t,\omega ))$ are also continuous functions and bounded functions for all $x\in \mathbb{R}$.

For transformation and simplicity purposes, we would represent the following functions as $\mathcal{U}=\mathrm{\Phi}(x(t,\omega ),t)$ and $\mathcal{V}=\frac{d\mathrm{\Phi}(x(t,\omega ),t)}{dx}$.

Given the following conditions:

- (i)
$U={U}_{1}\times {U}_{2}\in {\mathbb{R}}^{m}$ is a compact and bounded set.

- (ii)
$I\in \mathbb{R}$ is bounded.

- (iii)$H=[a,b]$, $a>0$ and $b>0$$F:U\times I\times H\to {\mathbb{R}}^{2}\left({\parallel \cdot \parallel}_{{\ell}_{1}}\text{norm}\right).$

In summary we have the following results.

**Theorem 4.3**

*Consider a special case for the equation*

*for a one*-

*dimensional problem and for*

*Then*, *assuming that* (A:1) *and* (A:2) *hold*, *at least one solution has been found to exist*.

The solution in (23) is not necessarily unique, and to attain uniqueness, more boundary conditions to the ODE must be given. For a one-dimensional problem at least one solution has been found to exist, and for $n\ge 2$ the equation is a PDE which is difficult to solve.

### 4.1 Iterative optimal control estimates

The expectation of the value function is driven by stochastic differential equation (19). The function $\rho (\mathrm{\Upsilon}|{x}_{j})$ in equation (25) is the probability density function of the transitions, and the function ${\tilde{Q}}_{j}$ will be defined later.

Certainly, we cannot surely know future paths and the future control values due to the presence of the noise to the problem. This does not mean we have to give up since future paths cannot be certainly known, therefore we may estimate future paths, hence future control values, in order to attain optimums as the controls are dependent on the path control.

for ${\delta}_{j}={t}_{j+1}-{t}_{j}$, which is the change in time *t*.

Hence we have the following results given as a lemma.

**Lemma 4.1**

*From both Theorem*4.2

*and Theorem*4.3,

*and assuming that the transitions are given by equation*(27),

*we give the iterative optimal controls as*

*and*

*for the estimated value function*

*where*

*and*

*Proof*From Theorem 4.2, suppose that the solution is given as an estimated iterative value function in equation (25). Consider the discrete paths of the optimal trajectory given as

Applying Theorem 4.2, we obtain the optimal iterative control estimates, which completes the proof. □

## 5 Conclusion and future work

The paper studied the nonlinear stochastic control problem of a zero sum differential game. The problem studied here is similar to the one in [15], the difference is that ours is a minimax problem and the state function is dependent on an additional random variable *ω*, which was suggested as future work. Equating the control weights to the variance of the Brownian noise played an important role towards the solution. A quasilinear PDE was attained due to the logarithmic transformation of the value function, though the equation was difficult to solve. However, the quasilinear differential equation was converted to an ODE by fixing the time variable *t*, which showed that at least a solution (saddle point) does exist. The uniqueness of the solution can be attained if more boundary conditions are set to the problem. The problem is path-dependent and Gaussian distributed with mean zero and variance *θ*. Therefore, the Gaussian distribution has played a vital role in obtaining the control path estimates as given in Lemma 4.1. We are of the view that nonlinear Feynmac-Kac formula can be of great help in attaining the solution for the nonlinear PDE [23]. Lastly, note that for $n\ge 2$, the problem is a nonlinear PDE, which still remains an open problem according to our understanding.

## Declarations

### Acknowledgements

We would like to take this opportunity to thank the unknown referees and the Chinese Scholarship Council for giving us an opportunity to further our studies on this research area. Our sincere acknowledgements are also extended to our supervisor, families, colleagues for moral support and guidance.

## Authors’ Affiliations

## References

- Hellerstein JL, Morrison V, Eilebrecht E: Applying control theory in the real world: experience with building a controller for the .NET thread pool.
*ACM SIGMETRICS Perform. Eval. Rev.*2010, 37: 38-42. 10.1145/1710115.1710123View ArticleGoogle Scholar - Kendrick D: Applications of control theory to macroeconomics.
*Ann. Econ. Soc. Meas.*1976, 5: 171-190.Google Scholar - Kilian C:
*Modern Control Technology*. Thomson Delmar Learning, Clifton Park; 2005.Google Scholar - Kappen, HJ: Stochastic optimal control theory. ICML, Helsinki, Radbound University, Nijmegen, Netherlands (2008)Google Scholar
- Mazliak, L: An introduction to probabilistic methods In stochastic control. Laboratory Probabilities, University of Paris, France (1996)Google Scholar
- Yong J, Zhou X:
*Stochastic Controls. Hamiltonian Systems and HJB Equations*. Springer, Berlin; 1999.Google Scholar - Vorob’ev NH:
*Game Theory*. Springer, Berlin; 1977.View ArticleGoogle Scholar - Tirole J, Fundenberg D:
*Game Theory*. MIT Press, Cambridge; 1991.Google Scholar - Basar, T: Lecture notes on non-cooperative game theory. University of Illinois, Urbana (2010)Google Scholar
- Ho YC, Bryson AE, Baron S: Differential games and optimal pursuit-evasion strategies.
*IEEE Trans. Autom. Control*1965, 10: 385-389. 10.1109/TAC.1965.1098197MathSciNetView ArticleGoogle Scholar - Yeung DWK: A feedback Nash equilibrium solution for noncooperative innovations in a stochastic differential game framework.
*Stoch. Anal. Appl.*1991, 9: 195-213. 10.1080/07362999108809234MathSciNetView ArticleGoogle Scholar - Browne S: Stochastic differential portfolio games.
*J. Appl. Probab.*2000, 37: 126-147.MathSciNetView ArticleGoogle Scholar - Theodorou, EA, Todorov, E: Stochastic optimal control for nonlinear Markov jump diffusion processes. PhD thesis, University of Southern Carlifornia (2011)Google Scholar
- Fleming WH, Soner HM:
*Controlled Markov Processes and Viscosity Solutions: Applications of Mathematics*. 2nd edition. Springer, New York; 2006.Google Scholar - Theodorou, EA: Iterative path integral stochastic optimal theory: theory and applications to motor control. Thesis, University of Southern California (2011)Google Scholar
- Guo X, Liu J, Zhou XY: A constrained non-linear regular-singular stochastic control problem, with applications.
*Stoch. Process. Appl.*2004, 109: 167-187. 10.1016/j.spa.2003.09.008MathSciNetView ArticleGoogle Scholar - Oksendal BK:
*Stochastic Differential Equations: An Introduction with Applications*. 6th edition. Springer, Berlin; 2003.View ArticleGoogle Scholar - Josa-Fombellida R, Rincon-Zapatero JP: New approach to stochastic optimal control.
*J. Optim. Theory Appl.*2007, 135: 163-177. 10.1007/s10957-007-9262-5MathSciNetView ArticleGoogle Scholar - Silva, F: Interior penalty approximation for optimal control problems. Optimality conditions in stochastic optimal control theory. CMAP, Ecole Polytechnique, and INRIA Saclay (2010)Google Scholar
- Mou L, Yong J: Two person zero sum linear quadratic stochastic differential games by Hilbert space method.
*J. Ind. Manag. Optim.*2006, 2(1):93-115.MathSciNetGoogle Scholar - Rigobon R:
*Brownian Motion and Stochastic Calculus Introductory Notes*. MIT Press, Cambridge; 2009.Google Scholar - Chow, GC: Optimal control without solving the Bellman equation. Econometric Research Program. Research Memo No. 364, Princeton University (1992)Google Scholar
- Peng, S, Wang, F: BSDE, path-dependent PDE and nonlinear Feynmac-Kac formula. School of Mathematics, Shandong University, Jinan, China (2011)Google Scholar

## Copyright

This article is published under license to BioMed Central Ltd. This is an Open Access article distributed under the terms of the Creative Commons Attribution License (http://creativecommons.org/licenses/by/4.0), which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly credited.